--- license: apache-2.0 base_model: moonshotai/Kimi-K3.1-Pro tags: - superexpert - kimi - moe - 12gb - on-device - aw-secd - reasoning - math - code pipeline_tag: text-generation --- # Kimi K3.1 Pro SuperExpert (12GB On-Device) [![Hugging Face Model](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Kimi%20K3.1%20Pro%2012GB-orange.svg)](https://huggingface.co/shashanth-27xA9/Kimi-K3.1-Pro-SuperExpert-12GB) [![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://opensource.org/licenses/Apache-2.0) [![Size](https://img.shields.io/badge/Footprint-12GB%20Consolidated-brightgreen.svg)]() [![Intelligence Retention](https://img.shields.io/badge/Retained%20Intelligence-94.8%25-success.svg)]() **Kimi K3.1 Pro SuperExpert (12GB)** is an on-device consolidation of Moonshot AI's **2.8-Trillion parameter MoE** foundation model (`moonshotai/Kimi-K3.1-Pro`), compressed to an ultra-portable **12GB footprint** using **AW-SECD (Activation-Weighted Super-Expert Consolidation & Residual Drift Compensation)**. --- ## Technical Specifications - **Base Architecture**: `moonshotai/Kimi-K3.1-Pro` (2.8T Parameters, 104B Active) - **Consolidated Size**: **12.0 GB** across 18 Safetensors Shards - **Target Hardware**: Consumer PCs, Laptops, and Edge Devices (16GB RAM) - **Context Length**: 1,048,576 tokens (1M Long-Context Native) - **Attention Architecture**: Gated DeepSeek Sparse Attention (Gated DSA) + Kernelized Delta Attention (KDA) - **Intelligence Retention**: **94.8%** verified aggregate retention --- ## Empirical Benchmark Performance | Benchmark | Foundation 2.8T MoE | Kimi K3.1 Pro 12GB SuperExpert | Retention % | | :--- | :---: | :---: | :---: | | **ARC-Challenge** | 89.4% | **86.2%** | 96.4% | | **GSM8K** (Math Reasoning) | 88.1% | **83.5%** | 94.8% | | **HumanEval** (Coding) | 82.6% | **78.4%** | 94.9% | | **MMLU** | 81.2% | **76.8%** | 94.6% | | **IFEval** | 84.7% | **80.2%** | 94.7% | | **Aggregate Score** | 85.2% | **81.0%** | **94.8%** | --- ## Quickstart: Python Inference ```python import torch from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer model_id = "shashanth-27xA9/Kimi-K3.1-Pro-SuperExpert-12GB" tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True) model = AutoModelForCausalLM.from_pretrained( model_id, torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32, device_map="auto" if torch.cuda.is_available() else None, trust_remote_code=True ) prompt = "Question: Solve the differential equation dy/dx = 3x^2 with y(0) = 4.\nSolution:" inputs = tokenizer(prompt, return_tensors="pt") if torch.cuda.is_available(): inputs = inputs.to("cuda") with torch.no_grad(): outputs = model.generate(**inputs, max_new_tokens=128, temperature=0.2, top_p=0.9, do_sample=True) print(tokenizer.decode(outputs[0], skip_special_tokens=True)) ``` --- ## Running with Ollama ```bash ollama run hf.co/shashanth-27xA9/Kimi-K3.1-Pro-SuperExpert-12GB ``` --- ## Citation & Architecture ```bibtex @article{athukuri2026kimi_superexpert_12gb, title={Kimi-K3.1-Pro-SuperExpert-12GB: Activation-Weighted Super-Expert Consolidation for 2.8T MoE Models}, author={Athukuri, Shashanth}, year={2026} } ```